Articles | Autumn 2006

COMPUTING POTENTIAL OUTPUT AND THE OUTPUT GAP

FOR THE PORTUGUESE ECONOMY*

Vanda Almeida**

Ricardo Félix**

1. INTRODUCTION

Gross Domestic Product (GDP) is one of the main welfare indicators in developed economies. It is cer-
tainly the most widely used indicator whenever the economic prosperity of a country is an issue to be
assessed. In the case of the Portuguese economy, this indicator points towards a weak economic per-
formance in the recent past, and deserves a deeper analysis that goes beyond the usual conjunctural
assessment of the evolution of each aggregate demand component. This article suggests an interpre-
tation of this phenomenon based on economic growth theory and on the concept of potential output,
enabling an assessment of the supply side conditions and the identification of some structural factors
that might have limited the Portuguese economy growth since the beginning of the millennium.

Using appropriate techniques, GDP can be decomposed into a structural component and a
conjunctural component. The first one is usually named as “potential output” and it can be defined as
“the level of output at which the economy’s resources are fully employed or, more realistically, at which
unemployment is at its natural rate” (Mankiw, 2003, pp. 246). The second component, usually named
“output gap”, is the deviation between the actual level of output and the potential output and it includes
temporary elements that are shaped by business cycle and other very short-run fluctuations.

The computation of potential output and the output gap for the Portuguese economy enables not only
an assessment of economic growth potential, but also the measurement of the cyclical position of the
economy and the identification of changes in the pattern of business cycle evolution. These indicators
usually play a relevant role in different domains of economic analysis, such as in the computation of
structural indicators (for instance, the cyclically-adjusted budget balance) and in the appraisal of infla-
tionary pressures in the economy stemming from the demand side. Additionally, these indicators are
also used in the assessment of the overall consistency of the macroeconomic projections for the
Portuguese economy.

Potential output is not an observable variable and must therefore be computed using an information
set that contains observable variables, using techniques that combine macroeconomic theory with sta-
tistics and econometrics. These techniques are usually classified into two broad categories: statistical
methods, which decompose mechanically real GDP time series into its trend, cycle and irregular com-
ponents; and structural methods, which use economic theory in the process of potential output compu-
tation. Since potential output resulting from the implementation of the previous methodologies is not an
observable variable, it is not possible to evaluate the accuracy of the computed figure based on the
usual goodness of fit measures, in contrast to what usually happens with observable variables. Thus,

* The opinions contained in this article belong to the authors and do not necessarily coincide with those of the Banco de Portugal. The authors would like to
acknowledge the useful comments of Nuno Alves, João Amador, António Antunes, Mário Centeno, Ana Cristina Leal and José Ferreira Machado, as well as
the cooperation of the Banco de Portugal Research Department Forecast Unit. Responsibility for any errors and omissions remains ours.
** Economic Research Department.

Economic Bulletin | Banco de Portugal 73

 Autumn 2006 | Articles

the point estimate obtained for each year in the sample should not be taken as the true figure, but as an
amount with a strong probability of being close to the true magnitude of the variable.

This article explores some of the most commonly used methods for computing potential output and the
output gap and applies them to Portuguese economy data. Among the statistical methods, the Hodrick
and Prescott (HP), the Baxter and King (BK) and the Christiano and Fitzgerald (CF) filtering methods
were implemented.1 Concerning the structural methods, the production function approach was consid-
ered in two alternative formulations: the CES2 production function and the Cobb-Douglas function,
which is a particular case of the former.

This article is organised as follows: section two describes the alternative methods used to compute po-
tential output; section three discusses the empirical results obtained by applying those methods; and
section four, presents the main conclusions of this study and suggests directions for further research.

2. METHODS FOR COMPUTING POTENTIAL OUTPUT

Potential output is a non-observable variable, as it measures a phenomenon that cannot be empirically

observed: the quantity of goods and services that an economy can produce by making full use of all its
available resources.

The need for computation of potential output has led to the development of several methodologies,
which combine distinct subjects of economic analysis such as macroeconomic theory, statistics and
macroeconometrics. These methods make use of information on observable variables to obtain re-
sults for potential output, and reflect not only the information they use but also the properties of the
techniques they apply. In general, these methods are grouped into statistical or structural methods, ac-
cording to the techniques they use and the information they incorporate.

2.1. Statistical methods

Statistical univariate methods3 are purely mechanical procedures, which can be used to decompose
any time series into different components. They are used for identifying components with strong per-
sistency, usually called tendency, and components with weak persistency, usually associated with cy-
clical and very short-run movements.

The direct application of these methods to the observed output series produces a smoothed series,
called trend output, which is taken to represent potential output. The remaining components (cyclical
and erratic) correspond to the output gap.

Univariate statistical methods have a direct interpretation and are generally easy to implement, being a
practical way to compute potential output. However, these techniques do not incorporate any macro-
economic theory, which limits the interpretation of the results they produce and restricts its utility in the
analysis of the behaviour of the economy. Additionally, statistical methods suffer, in general, from
end-of-sample problems, which are usually tackled by extending the output series, introducing, how-
ever, a new source of uncertainty in the computation process. Finally, it is also important to account for
problems in the treatment of structural breaks, since these methods spread the impact of a break
through many periods, influencing potential output growth in several periods instead of having effects
strictly at the moment in which the break occurred.

(1) For further details see Hodrick and Prescott (1997), Baxter and King (1999) and Christiano and Fitzgerald (1999).
(2) CES is the acronym for Constant Elasticity of Substitution.
(3) Statistical methods can be grouped into two categories, univariate and multivariate. In this article, we will focus on the univariate methodologies.

74 Banco de Portugal | Economic Bulletin

 Articles | Autumn 2006

In practice, statistical methods simply apply bilateral moving averages to the historical output series, to
filter out components with distinct frequencies. Among the several univariate techniques available, this
article covers three main methods, which are frequent in the literature on economic growth and busi-
ness cycles: the Hodrick-Prescott filter (HP) and the band-pass filters Baxter and King (BK) and
Christiano and Fitzgerald (CF).

The HP filter is a simple smoothing procedure that extracts a non-linear trend component from ob-
served output by minimising a weighted average of the variability in the trend and its deviations from
actual output. Formally, trend output is obtained through the minimization of the following loss function:

S S -1
minT L = å ( y t - y Tt )2 + l å ( Dy t + 1 - Dy t )2
T T

yt t =1 t=2 (1)

where y t and y Tt are the logs of observed and trend output respectively,4 S is the number of observa-
tions and l is a smoothing parameter. The minimization of the loss function implies a choice of a value
for the smoothing parameter, which represents a penalty for variability in the growth of trend output: a
higher value for l implies smoother estimates for trend output and consequently a smaller output gap.

The HP filter has some good features that have contributed to its wide utilisation, including the fact that
it renders the output gap stationary, as presented in King and Rebelo (1993), and the fact that it is flexi-
ble and simple to implement. However, there are also some problems associated with the HP filter. The
first one comes from the choice of the appropriate smoothing parameter l, which is largely discretion-
ary and far from being consensual. A second shortcoming is commonly known as the end-of-sample
problem, which results from the fact that towards the edges of the sample, as leads and lags become
unavailable, the HP filter gradually turns into an asymmetric filter, overemphasising the importance of
the last observations.5 This way, estimates of trend output for recent history suffer from bias, which is
particularly serious because estimates for recent periods are typically those in which policymakers are
more interested for purposes of policy decision. A common way of tackling this issue is to extend the
output series forward using reliable projections. Finally, many studies refer6 that when used with data
that is integrated or nearly integrated, the HP filter can induce spurious cycles, i.e. it can generate
cycles even if they are not present in the original data.

Another univariate type of filtering technique is the band-pass filter, which relies on the theory of spec-
tral analysis of time series data. This methodology is used to transpose time series fluctuations repre-
sented in the time domain, to fluctuations in the frequency domain. Assuming that business cycle
fluctuations correspond to a well-defined band of frequencies, it is then possible to apply the filter to the
observed output series, and isolate the observations corresponding to the pre-defined band, obtaining
the output’s cyclical component. In practice, the filtered series consists of a weighted average of the
original time series, where the weights attributed to each component of the series are determined ac-
cording to the frequencies we wish to retain. The filter is the array of weights that, when applied to the
original series, produces the cyclical component of output. More formally, the filtered series is given by:

¥
B(L) y t = åb L y j
j
t
j = -¥ (2)

where B(L) represents the band-pass filter, b j corresponds to the weight attributed to y t - j , and
L j stands for the usual lag operator.7

(4) Variables in small letters represent the natural logarithms of the correspondent variables in big letters.
(5) For further details see Giorno et al. (1995), Cerra and Saxena (2000) and Mohr (2005). 5
55555

(6) See, for example, Harvey and Jaeger (1993) or Cogley and Nason (1995).
(7) For any variable X t , the lag operator L j is defined such that L j X t = X t - j . 7
77777

Economic Bulletin | Banco de Portugal 75

 Autumn 2006 | Articles

The BK and CF filters are probably the two most widely used band-pass filters. The BK filter performs a
finite and symmetric bilateral moving average, imposing the same number of leads and lags and sym-
metry of weights, which implies that observations in a similar position on each side of the central obser-
vation are given equal weights. These characteristics have the advantage of assuring that the filtered
series has no phase shift, i.e., that the timing of peaks and troughs is consistent with the behaviour of
the unfiltered series. However, this is achieved at a cost of losing observations at the beginning and
end of the sample. Usually, to solve this problem, the series are extended using the same type of
techniques mentioned for the HP filter.

The CF filter, contrary to the BK filter, uses all observations available in the sample, forward and back-
wards, implying that in each period, the number of leads differs from the number of lags, which turns
the filter asymmetric with a varying weighting scheme. This way, the CF filter overcomes one of the lim-
itations of the BK filter, the loss of observations at the beginning and the end of the sample, but it may
introduce the phase-shift problem already explained.

2.2. Structural methods: the production function approach

Structural methods, contrary to statistical, take economic theory into account in the process of compu-
tation of potential output and the output gap. They establish a link between potential output and other
macroeconomic variables, which introduces the opportunity of examining the underlying economic
factors that drive changes in potential output and therefore gives the possibility of deriving some eco-
nomic interpretation from the evolution of the obtained results, instead of just assessing their value.
However, the application of this type of methods implies a prior choice of an adequate model, which
necessarily consists of a simplification of reality and relies on a set of assumptions about the structure
of the economy that may or may not be entirely correct. In addition, these methods require a large
amount of information, which may be a problem in situations where there are limitations in assessing
the data or when their quality is questionable. It is also worth noting that, in general, structural methods
are still dependent on the utilization of statistical univariate methods to calculate trend components for
some variables used in the computation process, which revives the limitations associated with
statistical methods, already discussed.

One of the keystones among structural methods is the production function approach. This approach
looks at the supply side conditions of the economy, postulating an aggregate production function that
explicitly models output as the outcome of a production process, depending on: the available quantity
of factors of production, the productivity of these factors, and their weight in output. Potential output is
obtained as the result of the defined production function when its contributing inputs and productivity
are at their sustainable long-run levels.

When compared to other structural methods, the production function approach has the important ad-
vantage of allowing for an explicit growth accounting exercise, which expresses potential output
growth as a function of the growth rate of each of its determinants.

2.2.1. The CD and CES production functions

The functional form adopted for the production function synthesizes in a very simple way the technol-
ogy used in the productive process, i.e. the way factor inputs are combined to produce output. There is
not, however, a consensus as to the best functional form to adopt, and several different forms have
been proposed in the literature. Among these, the two most widely used, in the literature on economic
growth, are the CES production function, and a particular case of it, the CD function. These functions

76 Banco de Portugal | Economic Bulletin

 Articles | Autumn 2006

differ in many aspects, including in the complexity of their functional form and in the restrictions they
impose on the technology underlying the production of goods and services in the economy.

The CD function is, most probably, the most widely used functional form for the production function,
mainly due to the fact that it is analytically simple and straightforward to calibrate. Furthermore, when
one looks at data concerning long periods of time, the characteristics of the CD function seem to be
compatible with the observed facts for a wide range of economies.

The most popular functional form, for the CD production function, considers two productive factors,
capital and labour, and is formally expressed as:

a 1- a
Yt = AL
t t Kt , 0 < a <1 (3)

where At represents total factor productivity (TFP), Kt corresponds to the capital stock, Lt is total em-
ployment, and a is a constant that corresponds to the elasticity of output with respect to labour. This
elasticity is calibrated to match the empirical average labour share obtained from the national accounts
data, being in line with one of the major assumptions of the CD function, the stability of the income fac-
tor shares in output. It is therefore important to keep in mind that in order to be able to apply the CD
function, the historical information on the income factor shares must point to their constancy over time.
Furthermore, this assumption has the important implication of unitary elasticity of substitution between
factors of production. This is a rather restrictive assumption, meaning that an increase in the relative
price of one of the factors will always be accompanied by a proportional decrease in the relative
utilisation of that factor.

A more general alternative to the CD functional form is provided by the CES production function. In this
article, we use the following specification:
s

Yt = é d (BtLt ) + (1- d)( X tKt ) s ù , 0 < d < 1 and s > 0

s- 1 s- 1
s- 1
s

êë úû (4)

where Bt and X t are indexes of labour and capital augmenting technical progress respectively, d is the
distribution parameter capturing the functional distribution of income and s is the elasticity of
substitution.

This framework implies that factor income shares vary proportionally to factors’ real cost and produc-
tivity, which contrasts with the constancy assumption underlying the CD function. Furthermore, the
CES specification introduces an important advantage over the CD form since it does not impose the
substitution parameter to be equal to one, and thus gives the possibility of estimating it from the data. It
can be shown that when the elasticity of substitution is unitary, the CES function converges to a CD
function.

We see that the CES production function is markedly less restrictive than the CD function, allowing for
much richer results, specifically to test the validity of the CD formulation through the realization of a sta-
tistical test on the estimate obtained for the elasticity of substitution. In addition, the CES formulation
adopted in this article includes technical progress specific to each factor,8 which is not possible with the
CD function, since only total factor productivity is identified.

(8) The utilization of a CES production function with technical progress for capital is compatible with a stationary model only if technical progress is itself
stationary. For further details see Barro and Sala-i-Martin (1995).

Economic Bulletin | Banco de Portugal 77

 Autumn 2006 | Articles

2.2.2. The computation of potential output and the growth accounting exercise

In order to compute potential output, it is necessary to know beforehand the levels of potential factor
utilization and productivity, and estimates of the parameters needed for the production function.

Regarding factors of production, the computation of their potential levels follows the same procedure
for both the CD and CES functions. For capital input, it is quite common to use the actual capital stock
to measure both actual and potential capital, since capital is a relatively fixed input, at least in the short
run. This assumption seems valid, as long as there is no noticeable deviation of the capital stock from
its long-run level. As for potential employment, it is generally obtained through the natural rate of un-
employment and active labour force. Thus, potential factor levels are given by:

Kt* = Kt (5)

(
L*t = PAt 1- u t* ) (6)

where PAt corresponds to the observed active labour force and u t* is a measure of the natural rate of
unemployment, with both variables being exogenous.

The parameter a of the CD function is calibrated using the average labour income share obtained from
the national accounts data. Knowing this parameter, the unobservable total factor productivity is ob-
tained by computing the Solow Residual by solving the production function to At :

Yt
At =
Lat K1t - a (7)

The resulting Solow residual is then smoothed9 to obtain an estimate of trend factor productivity, At*.
With potential capital input Kt*, potential labour input L*t , and trend TFP At* we can now easily compute
potential output by plugging these values into the production function.

( ) (K )
a 1- a
Yt * = At* L*t *
t (8)

This expression can be directly applied to the growth accounting exercise, since it is log-linear in the
factors of production. Differentiating both sides of the production function in logs yields:

D{
y t* = D{
a t* + a{ Dl t* + (1- a ) Dk t* (9)
Potential ouput TFP Labour
14243
growth rate contribution contribution Capital
contribution

In the case of the CES production function, the computation process is more complex, since there are
two unknown parameters and, using the formulation adopted in this article, there are two specific
productivities to calculate. Solving the profit maximization problem with CES technology, we obtain the
first order condition for labour demand, which is given by:

y t - l t = s(w t - p t ) + (1- s)b t - s ln d (10)

Equation (10) can be interpreted as a long-run relation between output per worker, y t - l t , the real cost
of labour, w t - p t and the labour-augmenting technological progress, b t . It is therefore possible to con-
sider this equation as a cointegration relation, and estimate the elasticity of substitution between fac-
tors using Johansen’s maximum likelihood method,10 which produces efficient estimates for the
parameters of a cointegration relation.

(9) In this article, as commonly accepted in the literature, we use the HP filter as a smoothing technique to obtain trend TFP.
(10) A detailed description of Johansen’s method can be found in Johansen (1995).

78 Banco de Portugal | Economic Bulletin

 Articles | Autumn 2006

To obtain the labour productivity index, we make use of a standard procedure in the literature,11 and as-
sume that it grows at a constant rate. This way, its level can be well approximated by a linear trend,
which converts equation (10) into:

y t - l t = s (w t - p t ) + (1- s) C + hLt - s ln d ( ) (11)

where hL is the average growth rate of labour productivity, t is a deterministic trend and C is an un-
known scale constant. Even though the elasticity of substitution has been estimated, the distribution
parameter d and the scale constant C are still unknown. However, using the fact that the CES function
corresponds to a CD function in the case of unitary elasticity of substitution, d can be calibrated using
the equivalent parameter in the CD function (i.e., labour income share in value added, a), which is in
line with what is usually done in the literature. Using the obtained estimates for the parameters d and s
it is then possible to use equation (10) to find the scale constant, C, and the labour productivity index
Bt , given by the following expression:
1 s

æ Y ö 1- s æ P ö 1- s
Bt = çç t ÷÷ çç d t ÷÷
è Lt ø è Wt ø (12)

Finally, the capital-augmenting technological progress can be recovered by solving the production
function to X t :
s

æ 1 æY ö ö s-1
s- 1 s- 1
s
d æ BtLt ö s

X t = çç ç t÷ - ç ÷ ÷
ç 1- d çè Kt ÷ø 1- d çè Kt ÷ø ÷÷
è ø (13)

The calculated Bt and X t series are then smoothed using an HP filter, to obtain their trend levels, Bt* and
X t*, which corresponds to the same procedure used to smooth the Solow residual in the CD case.

Having obtained potential factor levels, their productivities, and the production function parameters,
potential output can finally be calculated by directly substituting them in the production function:
s

é
( ) ( ) ù s-1
s- 1 s- 1

Yt * = ê d Bt*L*t s
+ (1- d) X t*Kt* s

úû
ë (14)

For the CES production function, the growth accounting exercise is more complex than in the CD case,
since the CES function is not log-linear. The approach adopted in this article is to take the log of the
production function and then linearise it around the previous period, applying a first order Taylor expan-
sion. After some algebraic manipulation we get:

D{
y t* = wLt Db t* + wLt D*tl t* + wKt Dx t* + wKt Dk t* (15)
123 123 123 123
Potential output Labour productivity Labour Capital productivity Capital
growth rate contribution contribution contribution contribution

where:

( ) (1- d)( X t -1*Kt -1* )

s- 1 s- 1

d Bt -1 Lt -1
* * s s

w =
L
and w = K

( ) + (1- d)( X ) ( ) ( )
t s- 1 s- 1 t s- 1 s-1

d Bt -1 L t -1 d Bt -1 L t -1 + (1- d) X t -1 Kt -1
* * s * * s * * s * * s
t -1 Kt -1

represent the weights of the labour force and capital stock in output, which vary with time. In the case of
s = 1, these weights are given by wLt = d and wKt = 1- d, which exactly correspond to the income shares
of each factor in value added, since d was calibrated using the labour share of the CD function.

(11) See, for example, Dimitz (2001) and Jalava (2005).

Economic Bulletin | Banco de Portugal 79

 Autumn 2006 | Articles

3. POTENTIAL OUTPUT AND THE OUTPUT GAP FOR THE PORTUGUESE

ECONOMY

Potential output and the output gap for the Portuguese economy can be computed using the methods
presented in Section 2 and the available data. The results obtained are extremely useful in the assess-
ment of the evolution of economic activity, enabling, in particular, the identification of supply side fac-
tors that lie behind the weak economic growth witnessed in Portugal in recent years.

3.1. The dataset for the Portuguese economy

The dataset used in this article is mostly taken from the “Quarterly Series for the Portuguese Economy:
1977-2005” published in the summer 2006 issue of the Banco de Portugal Economic Bulletin.12

The time series for compensation per worker was built using employees’ wage bills and assuming that
a self-employed worker earns on average 75 per cent of an employee. Concerning the natural rate of
unemployment, it was assumed that it has remained broadly unchanged throughout the sample period
at 5.5 per cent of the labour force;13 nevertheless, the recent increase in long-term unemployment may
raise some doubts on the maintenance of the estimated natural rate of unemployment. The capital
stock time series was built using the perpetual inventory method, assuming a slightly increasing depre-
ciation rate that captures the faster depreciation of some types of investment goods (in particular, elec-
tronic equipment and computer systems). The value-added time series as well as the respective
deflator were computed from GDP at market prices by subtracting indirect taxes. Finally, income per
unit of capital was obtained using the resource constraint and the previously mentioned time series:

t t - WL
PY
Rt = t t

Kt (16)

It should not be disregarded that this is a very inaccurate measure of income per unit of capital, since
both compensation per worker and capital stock lie on the assumptions previously referred and do not
correspond to effectively observed figures. Thus, all measurement errors related with both labour in-
come and capital stock will translate directly to measurement errors in the income per unit of capital
considered.

3.2. Statistical methods

The implementation of the univariate methods described in section 2.1 implies not only the choice of
the parameters’ values for each of the filters considered, but also an extension of the actual real GDP
data to avoid the end-of-sample problems previously mentioned. Thus, the actual real GDP time series
was extended up to 2010, using the Banco de Portugal projections published in the summer 2006 is-
sue of the Economic Bulletin for 2006 and 2007 and the average growth rate recorded in the period
1993-2005 for the rest of the extension period.

Concerning the HP filter, the smoothness parameter was set to l = 7680, which, according to Raven e
Uhlig (2002), corresponds to a smoothness parameter of 30 for annual data, corresponding to the

(12) This database corresponds to an update of the one published in Castro and Esteves (2004) and follows the methodology presented there.
(13) In line with the results published in Dias, Esteves and Félix (2004).

80 Banco de Portugal | Economic Bulletin

 Articles | Autumn 2006

benchmark usually considered in the Eurosystem exercises.14 In terms of the band-pass filters, a
low-pass specification was considered, removing all the fluctuations with a frequency lower than 12
years.

The annual average growth rates of the computed potential output using the HP, BK and CF filters for
the whole sample period as well as for the sub-samples are plotted in Charts 3.2.1 and 3.2.2. An imme-
diate conclusion is that the discrepancies resulting from alternative univariate methods (which are visi-
ble in Chart 3.2.2) tend to vanish whenever averages of sample periods are considered. The results
suggest an annual average growth rate at around 3 per cent for the whole sample period (1986-2005).
However, an inspection of sub-sample periods reveals that this annual average growth rate does not
result from a broadly flat profile, since annual average growth rates differ quite substantially across pe-
riods. In fact, the results suggest an annual average growth at around 4 per cent in the sample period
1985-1994 and only 2 per cent in the sample period 1995-2005, revealing a continued decline in the
potential output growth rate throughout the last 20 years. In particular, a closer inspection of the last 5
years of the sample suggests that the potential output annual average growth rate was probably not
more than 1.5 per cent.

The computed output gap is plotted in Chart 3.2.3. In general terms, one can easily conclude that de-
spite the fact that point estimates do not coincide, the computed output gap is broadly similar across
the alternative methods used and the turning points tend to coincide. The results suggest that by the
time of Portuguese accession to the European Union in 1986, output was significantly below its poten-
tial level. In the following years, real GDP growth surpassed potential output growth, determining a
computed output gap at around 4 per cent in 1990. Subsequently, the significant slowdown in real GDP
growth led to a decline in output gap that reached a close to zero position in 1993, declining further until
1995. In the period 1995-2001, the Portuguese economy returned to economic activity growth rates
above potential with the output gap reaching 3 per cent in 2001. Thereafter, the accumulation of
disequilibria with non negligible impact on aggregate demand level has limited real GDP to growth

Chart 3.2.1 Chart 3.2.2

POTENTIAL OUTPUT GROWTH – STATISTICAL POTENTIAL OUTPUT GROWTH – STATISTICAL

METHODS (SUB-SAMPLES) METHODS
Annual average growth rate, in percentage Annual average growth rate, in percentage

6% 8%
Observed GDP HP BK CF
HP BK CF
5%
6%

4%
4%

3%

2%
2%

0%
1%

0% -2%
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005

1986-2005 1986-1994 1995-2005 1995-2000 2001-2005

(14) As already referred in Section 2, the choice of the HP filter smoothing parameter is to a large extent discretionary. One must emphasize that the choice of
alternative smoothness parameters only affects the amplitude of the business cycle, while maintaining both the average potential output growth and the
turning points of each business cycle. In particular, the utilisation of the smoothness parameter value originally suggested in Hodrick and Prescott (1997),
l = 1600, leads to the computation of business cycles with a smaller amplitude and to a more volatile potential output growth rate. 14
14141414

Economic Bulletin | Banco de Portugal 81

 Autumn 2006 | Articles

Chart 3.2.3

OUTPUT GAP – STATISTICAL METHODS

As a percentage of potential output

6%
HP BK CF

4%

2%

0%

-2%

-4%

-6%
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005

rates below potential, determining the progressive closure of output gap until 2003 and its return to
negative grounds in subsequent years.

In conclusion, the results obtained by implementing the alternative univariate statistical methods are
identical and allow for a general characterisation of the trend component of the actual output time se-
ries. One may highlight the fact that, irrespective of the method, potential output growth of the Portu-
guese economy has decelerated substantially in the last 20 years and that the results point towards an
annual growth rate ranging from 1 to 1.5 per cent in the last 5 years.

3.3. Structural methods: the production function approach

The production function approach draws on the specification of a function that is compatible with the
most evident stylised facts of the economy. One of these facts, first suggested in Kaldor (1965), is that
the income share of each one of the productive factors in value added is broadly constant. The evolu-
tion of labour income share as a percentage of value added in the Portuguese economy since 1986 is
plotted in Chart 3.3.1. This information was used to calibrate the parameters a and d of the CD and
CES production functions, respectively, since they were set at the average labour income share for the
period 1992-2005 (64 per cent). It should be noted that while the labour income share fluctuated for the
period 1986-1992, since then it has remained broadly stable.

To evaluate the reasonability of the CD specification, a CES production function was considered and
the elasticity of substitution among factors was estimated using quarterly data for the sample period
1988-2005 based on the methodology presented in Section 2.2. The estimated elasticity of substitu-
tion is 0.65 and the corresponding standard deviation is 0.06, meaning that the null hypothesis of a unit
elasticity of substitution (in case of a unit elasticity of substitution the CES production function col-
lapses to CD production function) is rejected at a 1 per cent significance level.15 Thus, the results sug-
gest that the unit elasticity of substitution is not supported by empirical evidence for the sample period
considered.

(15) The result obtained for the elasticity of substitution is similar to the one published in Lucas (1990) for the US economy.

82 Banco de Portugal | Economic Bulletin

 Articles | Autumn 2006

Chart 3.3.1

LABOUR INCOME SHARE

As a percentage of value added

75%

70%

65%

60%

55%

50%
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005

To assess both the impact of using a CD production function (when a CES production function seems
to be, according with available evidence, the one that is more adequate) and the differences arising
from the utilisation of structural methods instead of univariate methods, potential output and the output
gap were computed using CD and CES production functions and compared with the results obtained
using the HP filter, which was used as the univariate benchmark method.16

The potential output growth rate and the corresponding output gap estimates computed using the HP
filter and the alternative production function specifications are plotted in the charts 3.3.2, 3.3.3 and
3.3.4. The most evident finding is the coincidence of the results obtained using the alternative produc-
tion function specifications. Additionally, the output gap results obtained using both the production
function approach and the univariate statistical methods are qualitatively identical, despite slightly
higher amplitude of the output gap when the HP filter is used.

Moreover, the results plotted in Chart 3.3.3 show that potential output growth computed using the pro-
duction function approach usually tends to be more volatile than the one computed using univariate
statistical methods. This feature results from the fact that the production function approach, due to its
structural nature, reflects not only trend productivity growth (which is necessarily smooth since it is ob-
tained from a univariate filtering procedure), but also the growth of the available production factors,
which is not necessarily smooth, reflecting supply side shocks, in particular in the labour force.

A comparison of the annual average growth rate of potential output for the whole sample period
(1985-2005) and for the sub-sample periods that were previously used in the case of univariate statisti-
cal methods is presented in Chart 3.3.2. The results obtained suggest that the computed annual aver-
age potential output growth is very similar across methods both for the whole sample period and for the
sub-samples.

To sum up, the results obtained using the production function approach seem to confirm the deceler-
ation of the potential output growth rate already shown by the univariate statistical methods. However,
in contrast with the statistical methods, structural methods provide some indication of the factors that
are likely to be behind potential output deceleration through the growth accounting exercise.

(16) The choice of the HP filter as the univariate benchmark method is due to the fact that it is a widely used method in the literature whenever potential output
and output gap figures are required.

Economic Bulletin | Banco de Portugal 83

 Autumn 2006 | Articles

Chart 3.3.2 Chart 3.3.3

POTENTIAL OUTPUT GROWTH – STRUCTURAL VS POTENTIAL OUTPUT GROWTH – STRUCTURAL VS

STATISTICAL METHODS (SUB-SAMPLES) STATISTICAL METHODS
Annual average growth rate, in percentage Annual average growth rate, in percentage

6% 8%
CD CES HP
Observed GDP CD CES HP
5% 6%

4%
4%

3%
2%

2%
0%

1%

-2%

1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
0%
1986-2005 1986-1994 1995-2005 1995-2000 2001-2005

Chart 3.3.4

OUTPUT GAP – STRUCTURAL VS STATISTICAL

METHODS
As a percentage of potential output

6%
CD CES HP

4%

2%

0%

-2%

-4%

-6%
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005

It should be noted that the CD production function only allows for identification of the contribution of
trend total factor productivity, while the CES production function formulation considered enables identi-
fication of the contribution of the trend productivity growth of each one of the production factors for po-
tential output growth. In addition, it must be referred that the current formulation implicitly considers a
constant capital stock utilisation rate17 and the maintenance of the number of hours per worker, mean-
ing that any decline in these variables determines an overestimation of the amount of factor services
effectively used and an underestimation of the specific productivity of the factor under consideration,
since it is obtained as a residual. In the case of capital services, this type of bias tends to be bounded
since capacity utilisation is likely to be a stationary variable; in the case of labour, however, the bias

(17) The available information on capacity utilisation for Portugal refers to the manufacturing sector and does not take into account, for instance, the services
sector, which accounts for a significant share in overall production.

84 Banco de Portugal | Economic Bulletin

 Articles | Autumn 2006

might be more significant since the number of hours per worker has declined since 1986. Neverthe-
less, according to the latest issue of the Employment and Labour Market Statistics published by the
OECD,18 the decline in the number of hours per worker is near 8 per cent in the period 1986-2004,
meaning that an eventual underestimation in the annual trend productivity growth should not exceed
0.5 per cent on annual average terms. Thus, the results obtained and their interpretation in qualitative
terms is likely to be robust to the measurement problems previously mentioned.

The contribution from production factors and associated productivities to potential output growth is
mapped in Charts 3.3.5 and 3.3.6. According to the results obtained, the contribution of each factor, as
well as the contribution of trend total factor productivity is similar irrespective of the production function
used and thus the growth accounting exercise is robust to the production function specification consid-
ered. This conclusion applies both for the whole sample and for each of the sub-samples.

A first conclusion that can be drawn from the results obtained for the period 1986-1994 is that the
growth rate of factor productivity and the increase in capital stock played a crucial role in the real con-
vergence process observed since the accession of Portugal to the European Union.19 The growth of
labour has played a limited role, since it depends to a large extent on the evolution of demographic
structure, which is characterized by the ageing process.

Secondly, comparing the period 1995-2005 with the period 1986-1994, the reduction in the contribu-
tion of productivity is crucial to understand the reasons behind the decline in potential output and the
same applies to capital stock, though to a lesser extent. The CES production function results, which al-
low for identification of the productivity contribution of each factor, suggest that this decline in produc-
tivity is common to both capital and labour factors.

Finally, decomposing the sample period 1995-2005 into two sub-samples (1995-2000 and
2001-2005), one can conclude that the decline in potential output growth that is estimated to have oc-
curred in the last years stems essentially from a smaller contribution of capital stock growth rate and
from associated productivity, despite the slight decline in productivity associated to the labour factor.

Chart 3.3.5 Chart 3.3.6

GROWTH ACCOUNTING GROWTH ACCOUNTING

COBB-DOUGLAS CES

4.5% 4.5%

4.0% 4.0%

3.5% 3.5%

3.0% 3.0%

2.5% 2.5%

2.0% 2.0%

1.5% 1.5%

1.0% 1.0%

0.5% 0.5%

0.0% 0.0%
1986-2005 1986-1994 1995-2005 1995-2000 2001-2005 1986-2005 1986-1994 1995-2005 1995-2000 2001-2005
-0.5% -0.5%

Labour supply Capital stock Total factor prod. Labour supply Capital stock Labour prod. Capital prod.

(18) OECD (2006), “Average annual hours actually worked per worker”, Employment and Labour Market Statistics, 2006, release 01. 18
181818

(19) This result is in line with the conclusions presented in Cavalcanti (2004).

Economic Bulletin | Banco de Portugal 85

 Autumn 2006 | Articles

The decline in the contribution of capital stock reflects a distinct behaviour of investment in each one of
the sub-samples. Thus, while in the period 1995-2000 investment grew at around 8 per cent, in annual
average terms, it recorded an annual average decline of 3 per cent in the period 2001-2005. Moreover,
the decline in productivity associated to the capital factor may also be related with the progressive ob-
solescence of the capital stock already installed, since the decline in investment might have also even-
tually limited the normal process of replacement of the capital stock that depreciated in the meantime.

The main conclusion that can be drawn from the structural approach is that the deceleration of poten-
tial output recorded in recent years essentially reflects the unfavourable behaviour of investment and
its role in the maintenance of efficient conditions of production.

4. CONCLUDING REMARKS

This article presents computations for potential output and the corresponding output gap for the Portu-
guese economy, using alternative methods.

The results obtained are robust to the methodology adopted and point towards a deceleration of poten-
tial output in Portugal throughout the last 20 years, from an annual growth rate of around 4 per cent, for
the period 1986-1994, to an annual growth rate close to 1.5 per cent, for the period 2001-2005. It is
worth mentioning that these results are similar to those that can be found in European Central Bank
and European Commission working papers.20

The implementation of a structural method like the production function approach enables one to go be-
yond the mere description of potential output deceleration by identifying structural factors that are driv-
ing this evolution and that are, ultimately, the genesis of the weak economic growth witnessed in
Portugal in recent years. Nevertheless, interpretation of the results must be cautious, since they rely
on a number of previously mentioned assumptions.

The results suggest that the deceleration of potential output throughout the last 20 years has been
largely determined by a decline in the contribution of capital stock and total factor productivity. In the
period 1986-2004, the strong growth in potential output benefited from a very peculiar juncture. As re-
ferred in Cavalcanti (2004), this period corresponded to Portuguese accession to the European Union,
which may have implied a number of important transformations in the economy, in particular access to
new markets and improved financing conditions for the business sector, which are likely to have signifi-
cantly influenced both the dynamics of investment and improvement in total factor productivity. In the
last years (2001-2005), the weak growth of potential output reflects a limited contribution of capital
stock, as a result of the continued decline in investment since the beginning of the millennium, as well
as the impact of investment in the maintenance of efficiency conditions at factor productivity level, in
particular in the case of capital.

This study leaves a number of open questions that deserve some future research not only to reach a
deeper understanding of the conclusions and results just presented, but also to test their validity as
new information becomes available. Firstly, it seems important to consider the possibility of revisiting
the results using reliable information on hours worked instead of the number of workers, since the re-
sults obtained for total factor productivity contribution using the production function approach might
change. Secondly, the production function approach can be extended in order to account for the possi-
bility of considering imported intermediate goods (for instance, imported energy goods) as an addi-
tional factor to evaluate the impact of shocks in the price of these imported goods on potential output

(20) The estimates published by the ECB for potential output growth can be found in Benalal et al. (2006), while the estimates published by the European
Commission can be found in Denis et al. (2006).
20
202020

86 Banco de Portugal | Economic Bulletin

 Articles | Autumn 2006

level. Finally, the utilisation of methods that are able to combine the production function approach with
other structural approaches (for instance using Okun’s law and/or the Phillips curve), through
multivariate methods will make it possible to test the robustness of the results and conclusions
contained in this study using a larger information set.

REFERENCES

Barro, Robert. J. and Xavier Xala-i-Martin (1995): Economic Growth, McGraw Hill International
Editions, Economic Series.

Baxter, Marianne and Robert King (1999): “Measuring Business Cycles: Approximate Band-Pass
Filters for Economic Time Series”, Review of Economics and Statistics (81)4, 573-593.

Benalal, Nicholai, Juan Luis Diaz del Hoyo, Beatrice Pierluigi and Nick Vidalis (2006): “”Output growth
differentials across the euro area countries: some stylised facts”, Occasional Paper Series, n.
45, European Central Bank.

Berndt, Ernst R. and Davis O. Wood (1975): “Engineering and Econometric Interpretations of
Energy-Capital Complementarity”, The American Economic Review vol. 69(3), pp.342-354.

Berndt, Ernst R. and Davis O. Wood (1975): “Technology, prices, and the derived demand for energy”,
The Review of Economics and Statistics vol. 57(3), pp.259-268.

Castro, Gabriela and Paulo Esteves (2004): “Quarterly series for the Portuguese economy:
1977-2003”, Economic Bulletin, Banco de Portugal, June 2004.

Cavalcanti, Tiago V. (2004): “Business cycle and level accounting: the case of Portugal”, in
Proceedings of the conference “Portuguese economic development in the European area”,
March 2004, Banco de Portugal.

Christiano, Lawrence and Terry Fitzgerald (1999): “The Band pass filter”, Working Paper 9906,
Federal Reserve Bank of Cleveland.

Cogley, Timothy and James M. Nason (1995): “Effects of the Hodrick-Prescott filter on trend and
difference stationary time series Implications for business cycle research”, Journal of
Economic Dynamics and Control, Elsevier, vol. 19(1-2), pages 253-278.

Denis, Cécile, Daniel Grenouilleau, Kieran Mc Morrow and Werner Röger (2006): “Calculating
potential growth rates and output gaps - A revised production function approach”, European
Economy, Economic Papers, n.247, March 2006, European Commission, Brussels.

Dias, Francisco, Paulo Esteves and Ricardo Félix (2004): “Revisiting the NAIRU estimates for the
Portuguese economy”, Economic Bulletin, Banco de Portugal, June 2004.

Dimitz, Maria A. (2001): “Output-gaps and technological progress in European Monetary Union”,
Discussion Paper 20-2001, Bank of Finland.

Harvey, A. C. and Jaeger, A. (1993): “Detrending, Stylized Facts and the Business Cycle”, Journal of
Applied Econometrics, John Wiley & Sons, Ltd., vol. 8(3), pages 231-47, July-Sept.

Hodrick, R. J. and E. C. Prescott (1997): “Postwar U.S. business cycles: An empirical investigation”,
Journal of Money Credit and Banking 29, 1-16.

Jalava, Jukka, Matti Pohjola, Antti Ripatti and Jouko Vilmunen (2005): “Biased Technical Change and
Capital-labour Substitution in Finland, 1902-2003”, EU KLEMS Working Paper Series 4, EU
KLEMS Project.

Economic Bulletin | Banco de Portugal 87

 Autumn 2006 | Articles

Johansen, Soren (1995): “Identifying restrictions of linear equations with applications to simultaneous
equations and cointegration”, Journal of Econometrics, Elsevier, vol. 69(1), pages 111-132,
September.

King, Robert G. and Sérgio T. Rebelo (1993): “Transitional Dynamics and Economic Growth in the
Neoclassical Model”, American Economic Review, American Economic Association, vol.
83(4), pages 908-31, September.

Lucas, Robert E. (1990): “Supply-Side Economics: An Analytical Review”, Oxford Economic Papers,
New Series, vol. 42(2), pages 293-316, April.

Mankiw, N. Gregory (2003): Macroeconomics, Worth Publishers, 5th edition.

Raven, Morten O. and Harald Uhlig (2002):”On adjusting the Hodrick-Prescott filter for the frequency
of observations”, Review of Economics and Statistics (84)2, May 2002, 371-380.

88 Banco de Portugal | Economic Bulletin